Finite-sample analysis of nonlinear stochastic approximation with applications in reinforcement learning

In this paper, we study the bias and high-order error bounds of the Linear Stochastic Approximation (LSA) algorithm with Polyak-Ruppert (PR) averaging under Markovian noise. We focus on the version of the algorithm with constant step size and propose a novel decomposition of the bias via a linearization technique. We analyze the structure of the bias and show that the leading-order term is linear in the step size and cannot be eliminated by PR averaging. To address this, we apply the Richardson-Romberg (RR) extrapolation procedure, which effectively cancels the leading bias term. We derive high-order moment bounds for the RR iterates and show that the leading error term aligns with the asymptotically optimal covariance matrix of the vanilla averaged LSA iterates. We validate applicability of our findings for the temporal difference algorithm in reinforcement learning.

Motivated by the recent success of reinforcement learning algorithms, this paper studies a class of biased stochastic approximation (SA) procedures under a mild "ergodicity-like" assumption on the random noise sequence. Building on a multistep Lyapunov function that looks ahead to several future updates to accommodate the stochastic perturbations (thus gaining control over the bias), we prove a general result on the convergence of the SA iterates, and use it to derive non-asymptotic bounds on the mean-square error in the case of constant stepsizes. This novel viewpoint renders finite-time analysis of biased SA algorithms under a family of stochastic perturbations possible. For direct comparison with prior work, we demonstrate these bounds by applying them to TD-learning with linear function approximation, under the Markov chain observation model. The resultant finite-time error bound for TD-learning is the first of its kind, in the sense that it holds i) for the unmodified versions (i.e., without any modification to the updates) using even nonlinear approximators; as well as for Markov chains ii) under sublinear mixing conditions and iii) starting from any initial distribution, at least one of which has to be violated for existing results to be applicable.

In reinforcement learning, when the state space is enormous or infinite, it is not feasible to find the exact value for each state in the memory. A common way to tackle this problem is to adopt linear value function approximation technique. In this paper, we review some commonly used linear value function approximation algorithms, which divided into two classes. The first class is the so-called on-policy algorithms that mainly include temporal difference and least-squares-based methods. The second is the off-policy algorithms that use two strategies to solve the exploration-exploitation problem. From the perspective of algorithm design, on-policy is simple, but it does not take both exploration and exploitation into consideration when collecting data, while the off-policy can well balance the relation between exploration and exploitation. In the end, we first present some experiments to demonstrate the usefulness of linear value function approximation technique in the continuous and high dimensional scenes, and then in-class and out-class experiments are respectively conducted to further illustrate the performance of different approximation algorithms.KeywordsReinforcement learningLinear value function approximationLeast-squares temporal difference

In this paper, we study the initial butter decision (IBD) problem for a class of periodic production lines and the stochastic approximation (SA) algorithm for the problem. We establish the generalized semi-Markov process (GSMP) model of production line and the formulation of the IBD problem that is a bi-objective stochastic optimization problem. Considering the high computation cost of objective functions under middle scale stations, we employ SA algorithm to solve the optimization problem. To improve the performance to search the solutions on the boundary of feasible decision set, we propose an modified SA algorithm, called density stochastic approximation (DSA). Numerical results show SA algorithm is effective for the problem and DSA outperform SA and other two well-known variants of SA on searching the solutions on the boundary.

In this paper, we provide a new algorithm for the problem of stochastic global optimization where only noisy versions of the objective function are available. The algorithm is inspired by the well known cross entropy (CE) method. The algorithm takes the shape of a multi-timescale stochastic approximation algorithm, where we reuse the previous samples based on discounted averaging, and hence it saves the overall computational and storage cost. We provide proof of the stability and the global optimization property of our algorithm. The algorithm shows good performance on the noisy versions of global optimization benchmarks and outperforms a state-of-the-art algorithm for non-linear function approximation in reinforcement learning.

Many real-world tasks on practical control systems involve the learning and decision-making of multiple agents, under limited communications and observations. In this paper, we study the problem of networked multi-agent reinforcement learning (MARL), where multiple agents perform reinforcement learning in a common environment, and are able to exchange information via a possibly time-varying communication network. In particular, we focus on a collaborative MARL setting where each agent has individual reward functions, and the objective of all the agents is to maximize the network-wide averaged long-term return. To this end, we propose a fully decentralized actor-critic algorithm that only relies on neighbor-to-neighbor communications among agents. To promote the use of the algorithm on practical control systems, we focus on the setting with continuous state and action spaces, and adopt the newly proposed expected policy gradient to reduce the variance of the gradient estimate. We provide convergence guarantees for the algorithm when linear function approximation is employed, and corroborate our theoretical results via simulations.

This paper revisits the temporal difference (TD) learning algorithm for the policy evaluation tasks in reinforcement learning. Typically, the performance of TD(0) and TD( λ) is very sensitive to the choice of stepsizes. Oftentimes, TD(0) suffers from slow convergence. Motivated by the tight link between the TD(0) learning algorithm and the stochastic gradient methods, we develop a provably convergent adaptive projected variant of the TD(0) learning algorithm with linear function approximation that we term AdaTD(0). In contrast to the TD(0), AdaTD(0) is robust or less sensitive to the choice of stepsizes. Analytically, we establish that to reach an ϵ accuracy, the number of iterations needed is [Formula: see text] in the general case, where ρ represents the speed of the underlying Markov chain converges to the stationary distribution. This implies that the iteration complexity of AdaTD(0) is no worse than that of TD(0) in the worst case. When the stochastic semi-gradients are sparse, we provide theoretical acceleration of AdaTD(0). Going beyond TD(0), we develop an adaptive variant of TD( λ), which is referred to as AdaTD( λ). Empirically, we evaluate the performance of AdaTD(0) and AdaTD( λ) on several standard reinforcement learning tasks, which demonstrate the effectiveness of our new approaches.

Introduction Reinforcement learning formalizes the concept of learning from interactions.1 Broadly, reinforcement learning focuses on a setting in which an agent (decision maker) sequentially interacts with an environment that is partially unknown to them. At each stage, the agent takes an action and receives a reward. The objective of the agent is to maximize rewards accumulated in the long run. There are many situations in health care where decisions are made sequentially for which reinforcement learning approaches could prove useful for decision making. Throughout this article, we consider treatment prescription as an archetypical example to connect reinforcement learning concepts to a health care setting. In this setting, the care provider, the prescribed treatment, and the patients can be viewed as the agent, the action, and the environment, respectively, as depicted in Figure 1.Figure 1: Sequential treatment of AKI or CKD complications modeled as a reinforcement learning problem.Background In this section, with the objective of making reinforcement learning literature more accessible to a clinical audience, we briefly introduce related fundamental concepts and approaches. We refer the interested reader to Sutton and Barto1 for a comprehensive introduction to reinforcement learning. Markov Decision Processes Markov decision processes (MDPs) are a formalism of the sequential decision-making problem that has been central to the theoretical and practical advancements of reinforcement learning. In each stage of an MDP, the agent observes the state of the environment and takes an action, which, in turn, results in a change of the state. This change of state is assumed to be probabilistic with the next state being determined only by the preceding state, the chosen action, and the transition probability. The agent also receives a reward that is a function of the taken action, the preceding state, and the subsequent state. In an MDP, the objective of the agent is to maximize the return defined as the reward accumulated over a time horizon. In some applications, it is common to consider the horizon to be infinite, in which case the future rewards are discounted by a factor smaller than one. The selection of action by the agent on the basis of the observed state is known as the policy. More formally, a policy is a probabilistic mapping from states to each possible action. Because the policy and the reward are a function of the state, it is critical to estimate the utility of being in a certain state. More specifically, the value function is defined as the expected return starting from a given state under the chosen policy. Under this formalism, the objective of the agent is to find the optimal policy that maximizes the value function for all states. Reinforcement Learning Methods Action-value methods are a class of reinforcement learning methods in which the actions are chosen on the basis of the estimation of their long-term value. A prominent example of an action-value method is Q-learning in which the agent iteratively takes actions with the highest estimated values and updates the action-state value function on the basis of new observations. Policy gradient methods are another class of reinforcement learning methods that seek to optimize the policy directly instead of choosing actions on the basis of their respective estimated value. Such methods could be advantageous in health care applications that entail a large number of possible actions, e.g., when recommending a wide range of drug dosages or treatment options. Clinical Applications Reinforcement learning frameworks and methods are broadly applicable to clinical settings in which decisions are made sequentially. A prominent clinical application of reinforcement learning is for treatment recommendation, which has been studied across a variety of diseases and treatments including radiation and chemotherapy for cancer, brain stimulation for epilepsy, and treatment strategies for sepsis.2–5 In such treatment recommendation settings, a policy is commonly known as a dynamic treatment regime. There are various other clinical applications of reinforcement learning including diagnosis, medical imaging, and decision support tools (see refs. 2–5 and the references therein). Reinforcement Learning in Nephrology Although there have been recent applications of machine learning in nephrology,6,7 to the best of the authors' knowledge, the application of reinforcement learning to nephrology has been primarily limited to optimizing the erythropoietin dosage in hemodialysis patients.8,9 However, there are other settings where reinforcement learning has the potential to improve patient care in nephrology. For example, reinforcement learning methods can be adopted in the treatment of the complications of AKI or CKD (Figure 1). In this problem, the state models the conditions of the patient (e.g., vital signs, laboratory test results including urine and blood tests, and urine output measurements). The action refers to the treatment options (e.g., the dosage of medications such as sodium polystyrene sulfonate, and hemodialysis). The reward models the improvement in patient conditions. Similarly, reinforcement learning can help automate and optimize the dosage of immunosuppressive drugs in kidney transplants. Challenges and Opportunities Despite the success of reinforcement learning in several simplified clinical settings, their large-scale application to patient care faces several open challenges. The complexity of human biology complicates modeling clinical decision making as a reinforcement learning problem. The state space in such settings is often enormous, which could make a purely computational approach infeasible. Moreover, modeling all potential objectives a priori as a reward function may not be feasible. To overcome these challenges and realize the potential of reinforcement learning, clinical insight can play a pivotal role. More specifically, restricting the state space to only include highly relevant clinical variables could greatly reduce the computational complexity. Furthermore, using inverse reinforcement learning,2 relevant reward functions can be learned from retrospective studies assuming the optimality of clinical decisions. Another critical challenge is addressing moral and ethical concerns. It is imperative to ensure that reinforcement learning methods do not cause harm to the patient. To this end, there exists a need for a thorough validation of such methods before their use in patient care. Hence, there is a need to go beyond retrospective studies that have been used for the proof of concept of most existing reinforcement learning methods in health care applications.2,3 The lessons learned from the success of reinforcement learning in other application areas (e.g., self-driving cars) can help navigate the path to realizing its potential in health care. Accessible open-source simulation environments that enable researchers to compare various approaches are essential to the field of reinforcement learning. OpenAI Gym is currently the leading toolkit containing a wide range of simulated environments, e.g., surgical robotics.10 The development of high-quality and reliable simulation environments for nephrology and other health care applications can facilitate the development and validation of reinforcement learning methods beyond limited retrospective studies. The adoption of methods validated in such simulation environments in actual clinical settings will require clinicians' oversight. Similar to how self-driving cars require a human driver to ensure collision avoidance, clinicians' oversight is critical to ensure the safety of the patients, especially in the early stages of the adoption of reinforcement learning methods. The data from clinicians' decisions (e.g., overruling the automated treatment recommendation) can be used to improve the reliability of autonomous systems over time and reduce the burden of clinicians' oversight.

Causal Reinforcement Learning: An Instrumental Variable Approach

In this work, we study the logarithmic regret for reinforcement learning (RL) with linear function approximation and adversarial corruptions, in the formulation of linear Markov decision processes (MDPs). Specifically, we consider the case where there exist adversarial corruptions over the reward functions, and the total amount of the corruptions of each step h across all episodes K is bounded by a corruption level C ≥ 0. We propose an algorithm, double-weighted least-squares value iteration with UCB (DW-LSVI-UCB), which leverages weighted linear regressions to learn the (corrupted) unknown reward parameters and unknown transition parameters simultaneously. We prove that DW-LSVI-UCB attains an O( d2H4 log2(1+K/δ) gapmin + CdH2) regret (omitting the dependence on lower order terms), where d is the ambient dimension of the feature mapping, H is the horizon length, gapmin is the minimal sub-optimality gap, and K is the number of episodes. Additionally, when there are no adversarial corruptions over reward functions, the regret of our algorithm improves the previous best result by an O(dH/ log K) factor.

The goal of this paper is to investigate linear approximations of random functions and permutations. Our motivation is twofold. First, before the distinguishability of a practical cipher from an ideal one can be analysed, the cryptanalyst must have an accurate understanding of the statistical behaviour of the ideal cipher. Secondly, this issue has been neglected both in old and in more recent studies, particularly when multiple linear approximations are being used simultaneously. Traditional models have been based on the average behaviour and simplified using other assumptions such as independence of the linear approximations. Multidimensional cryptanalysis was introduced to avoid making artificial assumptions about statistical independence of linear approximations. On the other hand, it has the drawback of including many trivial approximations that do not contribute to the attack but just cause a waste of time and memory. We show for the first time in this paper that the trivial approximations reduce the degree of freedom of the related χ2 distribution. Previously, the affine linear cryptanalysis was proposed to allow removing trivial approximations and, at the same time, admitting a solid statistical model. In this paper, we identify another type of multidimensional linear approximation, called Davies-Meyer approximation, which has similar advantages, and present full statistical models for both the affine and the Davies-Meyer type of multidimensional linear approximations. The new models given in this paper are realistic, accurate and easy to use. They are backed up by standard statistical tools such as Pearson’s χ2 test and finite population correction and demonstrated to work accurately using practical examples.

This paper presents an approach to combinatorial optimization problems using linear function approximation (LFA) to solve the Travelling Salesman Problem (TSP). We create a reinforcement learning model in which we parameterize our policy using linear function approximation instead of the more commonly used neural networks. We then evaluated our models based on two factors: training time and optimality. When we compared our results with a state-of-the-art neural network solver, we found that our model was able to solve the TSP accurately while using drastically less computational resources and time to train than the neural network algorithm (Kool et al., 2019).

A promising solution for achieving autonomous wireless sensor networks is to enable each node to harvest energy in its environment. To address the time-varying behavior of energy sources, each node embeds an energy manager responsible for dynamically adapting the power consumption of the node in order to maximize the quality of service while avoiding power failures. A novel energy management algorithm based on reinforcement learning (RLMan) is proposed in this paper. By continuously exploring the environment, RLMan adapts its energy management policy to time-varying environment, regarding both the harvested energy and the energy consumption of the node. Linear function approximations are used to achieve very low computational and memory footprint, making RLMan suitable for resource-constrained systems, such as wireless sensor nodes. Moreover, RLMan only requires the state of charge of the energy storage device to operate, which makes it practical to implement. Exhaustive simulations using real measurements of indoor light and outdoor wind show that RLMan outperforms current state-of-the-art approaches, by enabling almost 70% gain regarding the average packet rate. Moreover, RLMan is more robust to variability of the node energy consumption.

Reinforcement learning is challenging if state and action spaces are continuous. The discretization of state and action spaces and real-time adaptation of the discretization are critical issues in reinforcement learning problems.In our contribution we consider the adaptive discretization, and introduce a sparse gradient-based direct policy search method. We address the issue of efficient states/actions selection in the gradient-based direct policy search based on imposing sparsity through the L 1 penalty term. We propose to start learning with a fine discretization of state space and to induce sparsity via the L 1 norm.We compare the proposed approach to state-of-the art methods, such as progressive widening Q-learning which updates the discretization of the states adaptively, and to classic as well as sparse Q-learning with linear function approximation. We demonstrate by our experiments on standard reinforcement learning challenges that the proposed approach is efficient.KeywordsDirect policy search Q-learningmodel selection

Energy harvesting (EH) two-hop communications are considered. The transmitter and the relay harvest energy from the environment and use it exclusively for transmitting data. A data arrival process is assumed at the transmitter. At the relay, a finite data buffer is used to store the received data. We consider a realistic scenario in which the EH nodes have only local causal information, i.e., at any time instant, each EH node only knows the current value of its EH process, channel state and data arrival process. Our goal is to find a power allocation policy to maximize the throughput at the receiver. We show that because the EH nodes have local causal information, the two-hop communication problem can be separated into two point-to-point problems. Consequently, independent power allocation problems are solved at each EH node. To find the power allocation policy, reinforcement learning with linear function approximation is applied. Moreover, to perform function approximation two feature functions which consider the data arrival process are introduced. Numerical results show that the proposed approach has only a small degradation as compared to the offline optimum case. Furthermore, we show that with the use of the proposed feature functions a better performance is achieved compared to standard approximation techniques.